Evolutionary Feature Combination Based Seed Learning for Diffusion-Based Saliency

Abstract

Diffusion-based saliency detection is a graph-based technique in which the optimal saliency map is computed by saliency propagation over the graph using diffusion of saliency values from one node to another. This is achieved by computing the product of a propagation matrix and a saliency seed vector. The saliency seeds stored in the saliency seed vector contain important prior saliency information usually obtained from a bottom-up saliency model or certain heuristics. Finding the optimal saliency seeds is vital for efficient saliency propagation during the diffusion process. In this work, we propose to investigate the performance of an evolutionary feature combination technique for learning the optimal seeds for diffusion-based saliency detection. We achieve this by adapting an evolutionary feature combination system (having good object detection performance) for the task of seed generation, for diffusion-based saliency, termed as IGASeed. We present quantitative and qualitative comparison of our proposed IGASeed system with the state-of-the-art heuristic and learning approaches for seed prediction. Our results show that our IGASeed technique performs better than most state-of-the-art models and comparable to the best seed learning model with lower computational cost.